Is AMD prepping a new R9 380X to tackle the midrange GPU market?

Ever since AMD debuted the R9 285 more than a year ago, there’s been speculation about whether or not the chip, codenamed Tonga, was a “full” implementation. Its die size and power consumption have always hinted that there might be more GPU horsepower under the hood, especially after AMD acknowledged that the chip had an additional four compute units that weren’t enabled on the original card. Now, “Fat Tonga” has supposedly been spotted in the wild, courtesy of Expreview in Japan and early shots of an XFX GPU.

If accurate, it implies that AMD has one more refreshed part to launch. Currently, the R9 380 (1792 cores, 112 TMUs, 32 ROPS) is a $199 part with 2GB and 4GB of RAM, while the R9 390 is a 2560:160:64 GPU with 8GB of RAM and a $329(Kshs.34,900) price tag. It’s not hard to see where a refreshed R9 380X would fit. A 2048:128:32 configuration would combine the original HD 7970 / R9 280X’s GPU firepower with Tonga’s bandwidth compression and slightly improved performance-per-watt. The major questions are how large the memory interface would be (AMD has only admitted to a 256-bit interface on Tonga, but 384-bit has always been rumored) and how AMD would set the clock speeds. AMD could theoretically drop 6-8GB of RAM on the card if it wanted to offer a unique option, but since the price point is likely below $300(Kshs.31,800), we’re betting on a 4GB SKU.

Our hypothetical R9 380X should be moderately faster than the old R9 285 or current R9 380, but it’s unlikely to break speed records or blow anyone’s doors off. At $249(Kshs.26,400), the most likely price, the R9 380X would be a bit faster than the GTX 960 and R9 380, which are generally evenly matched. It’s still well below the ~$329 level of the R9 390 or GTX 970, however, and of course these prices can fluctuate depending on which sales and specials are running at various retailers.

Once the R9 380X launches (assuming that’s what this is), AMD’s 300 series is likely complete. A number of you have asked if AMD will bring HBM lower in the product stack this year, but the answer to that is probably not. HBM and GDDR5 require different memory controllers, and AMD would have to rework its chips to make that happen, without any guarantee of higher sales. With 14/16nm GPUs expected next year, the benefits of launching HBM1 parts now as opposed to waiting for HBM2 in a year or less are very small.
If AMD has held on to the 384-bit 4-6GB version of Tonga for this long, it’s likely because the company wanted one more ace up its sleeve to respond with. Hopefully in 12 months we’ll be saying hello to GCN 2.0, as opposed to 1.3.